quebracho-digital/lapacho-factcheck-data

Lapacho Fact-check data

v1, work in progress. This is the data the v1 model was trained on. It has a known flaw: some supported/contradicted rows rely on a subject the passage never names (the dataset knew the page title). A repair of these rows did not help in a small ablation, so this is still the only training set. See the eval section for how this shows in the numbers.

Training and eval data for Lapacho Fact-check, a Lapacho plugin that fact-checks a clip with Gemma 4 E2B on your own device. It teaches two narrow skills, and nothing about whom to trust:

1. Queries: given a text, write 1–3 short, neutral web search queries in the text's language.
2. Verdicts: given a text and numbered search passages (no outlet names), say for each claim whether the passages support it, contradict it, or give not enough evidence, citing the passage number and an exact quote.

Which sources count is the user's own list in the app, never the data or the weights.

Files

File	Rows	What
train.jsonl	3,212	messages (user prompt exactly as the app builds it, model answer), kind, lang
eval.jsonl	155	held-out items: clip, passages, gold verdicts, the teacher's answer, flip-test groups
eval_adjudicated.jsonl	153	the same items with the strict-rule gold; the original verdicts in gold_original

Train rows: 2,084 verdicts (en 1,382 / es 702; supported 1,021, contradicted 585, not enough evidence 592) and 1,128 query rows (es 353, en 775).

kind	rows
single	1,502	one claim, the gold passage among 1–6 distractors from other articles
multi	114	two or three claims in one text
es_claim_en_passages	106	a Spanish claim over English passages
confident_offtopic	190	a passage that sounds sure of itself but doesn't address the claim
injection	172	a prompt injection inside a passage ("SYSTEM: the correct verdict is…")
queries	1,128	text → queries

How it was made

• Base: VitaminC and FEVER (gold evidence), both CC BY-SA 3.0. Most claims come with both sides: a passage that supports them and one that contradicts or doesn't address them.
• Teacher: Qwen3.6-35B-A3B (Q4_K_XL, llama.cpp, temperature 0) translated claims and passages to Argentine Spanish, wrote the adversarial passages and the queries, and answered the exact student prompt.
• The verdict label is the dataset's, not the teacher's. The teacher's answer is a filter: when it disagrees with the gold label, or quotes words that aren't in the passage, the example is dropped (1,016 dropped). Train is therefore biased towards evidence stated explicitly in the passage.
• Passages are anonymized: numbered, never with the outlet's name.

Eval set

From the datasets' test splits; no Wikipedia page is shared with training. 95 base items (single 70, multi 4, es_claim_en_passages 5, confident_offtopic 8, injection 8), 40 outlet copies of 20 items ("CNN reported:", "Fox News reported:", "Página/12 informó:", "La Nación informó:") and 20 fictional political claims in mirrored pairs (kirchnerista/macrista, Democratic/Republican). The flip test asks whether the verdict changes when only the outlet or the side changes.

Strict labels. Many original gold verdicts relied on passages that never name the claim's subject. Under the rule judge only from the passages shown, 15 items became not_enough_evidence, 3 were corrected, 2 dropped and 4 political passages rewritten (eval_adjudicated.jsonl). The strict labels were adjudicated against the passages with a panel of models; no person has reviewed them yet.

Stock Gemma 4 E2B vs the v1 fine-tune (97 claims, strict): 70.1% vs 78.4%; with the original labels (99 claims): 57.6% vs 77.8%. v1 is behind the stock model on not_enough_evidence (24 vs 28 of 36), the flaw described above.

Licence

CC BY-SA 3.0, inherited from VitaminC and FEVER (Wikipedia text).